Good Morning!
An AI named Tofu can now handle phone conversations. The future of customer service sounds increasingly like a grocery list.
Model | tokens | WoW (11/8) | |
1 | DeepSeek V4.1 Flash · DeepSeek | 9.11T | – +13% |
2 | Hy4 · Tencent | 1.93T | – -30% |
3 | GLM 5.3 Flash · Z.ai | 1.18T | – +4% |
4 | GPT-6 Luna · OpenAI | 0.93T | ▲2 +23% |
5 | Claude Opus 5.5 · Anthropic | 0.69T | ▲3 +0% |
6 | DeepSeek V4 Flash · DeepSeek | 0.61T | ▲1 -12% |
7 | MiniMax M3.1 Flash · MiniMax | 0.54T | ▲2 -0% |
8 | ernie-5.5-flash-preview · baidu | 0.53T | NEW |
9 | MiMo-V2.6-Flash · Xiaomi | 0.50T | ▲7 +61% |
10 | Nemotron 3.5 Super VL · NVIDIA | 0.48T | ▲3 +14% |
World Labs Opens Atlas API

TL;DR: World Labs opened Atlas to the public through browser apps and a developer API. Viewfinder reconstructs explorable 3D scenes from images, while Chisel lets creators guide the layout with simple geometry. The tools can turn Gaussian splats into meshes for use in games, simulations and 3D editors. Read more →
LongCat 2.5 on 194 Ascend Chips

TL;DR: Meituan’s LongCat-2.5 serving guide shows a 194-accelerator Huawei Atlas A2 setup: two for image encoding, 64 for prefill and 128 for decoding. In Meituan’s stress tests, running the model’s experts in 4-bit precision raised prefill throughput by 43.3% and decode throughput by 42.8% versus its 8-bit baseline. Read more →
Bland Previews Tofu-1 for Phone Calls

TL;DR: Bland opened research preview sign-ups for Tofu-1, a speech-to-speech model that listens and talks at the same time. It is designed to distinguish a caller from other voices on the line and handle corrections mid-sentence. In Bland’s demo, the agent keeps talking while it looks up information. Read more →
Haru Opens Full-Duplex Voice Weights

TL;DR: Higurashi released Haru-4B, a multilingual speech model that listens and speaks at the same time. It processes audio in 133 ms frames, supports nine languages and takes a short clip as a voice reference. The Apache 2.0 base weights come with PyTorch and Apple MLX inference code. Read more →
LongCat Extends Avatar Videos

TL;DR: Meituan released MIT-licensed weights for LongCat-Video-Avatar 1.6, which generates videos of people from audio. The team says the update improves lip sync and reduces visual drift in clips longer than 10 minutes. It also ships an optional INT8 checkpoint that stores the model’s diffusion weights at lower precision. Read more →
GPT-6 Astra Helps Develop a Physics Proof

TL;DR: Physicist Nathaniel Craig described how GPT-6 Astra helped him develop a candidate definition of the Standard Model on a spacetime lattice. Craig proposed the strategy, and Astra worked out how to refine the grid and led the detailed proof. The construction appeared in an October preprint, and outside experts reviewed the proof with Craig. Read more →
X3Guard Opens 150M Safety Weights

TL;DR: Axiomic Labs released X3Guard-150M, a small model that scores user prompts separately for harmful content and adversarial attempts such as jailbreaks or prompt injection. The Apache 2.0 weights come with three sensitivity presets. The team also published quantized ONNX versions for CPU inference and a browser demo. Read more →

Top AI models. Lower costs.
Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.